GPU solver that matches Fujitsu's $1M Digital Annealer on cubic optimization

1 comments

• bobchen1322 10 hours ago

    I built a CUDA solver for HUBO (Higher-Order Binary Optimization) — problems with 3+ variable interactions. Most solvers only handle
    pairwise (QUBO) and must "quadratize" cubic terms, inflating problem size 2-5x and degrading solutions.
  
    This solver operates on cubic terms natively with O(1) delta evaluation per bit-flip on GPU.
  
    Results on published benchmarks:
  
    - Cubic Knapsack (52 instances): 52/52 matched — ties Fujitsu Digital Annealer. 8.6 min total on a single RTX 3060 Ti.
    - Cubic Portfolio (n up to 1000): matches a dedicated solver; quadratized SA/Tabu drastically worse at every scale.
    - QUBO/MaxCut/MIS: competitive on standard quadratic benchmarks.
  
    Hardware: one RTX 3060 Ti (~$400). The comparison target (Fujitsu DA) is commercial hardware at ~$1M.
  
    If you have a combinatorial optimization problem — especially with higher-order terms — I'll benchmark it for free.